Modeling risk of stroke using binary logistic regression and multivariate adaptive regression splines

Stroke is a major source of disability and a major contributor to lost disability-adjusted life years. The study aims to compare Binary Logistic Regression (BLR) and Multivariate Adaptive Regression Spline (MARS) for modeling the risk of stroke. We use a dataset from IFLS (Indonesia Family Life Survey) and choose 800 samples randomly consisting of 200 stroke patients and 600 non-stroke patients. We divide the dataset into two parts for each validation, 75% for training and 25% for testing. Based on Chi-square test, all independent variables in this study, namely obesity, Diabetes Mellitus (DM), hypertension, and smoking status had a significant relationship with the risk of stroke (P-value<0.05). By using BLR and MARS, we find that the most significant variable which affect the risk of stroke is smoking status. In MARS methods, also find the interaction between predictors variable that can not be addressed by BLR. The results of this study show that the MARS approach gives better performance than binary logistic regression approach in estimating stroke risk with the accuracy obtained from training and testing data were 93.5%, while binary logistic regression approach gives accuracy of 91.3% in training data and 89% in the testing data.